Foreign Institutional Perspectives on TMT: Analysis of NVIDIA, Rubin Rack BOM, Semiconductors, and CPO - May 26, 2021

Key global TMT themes from foreign research reports on Thursday, May 26, 2021, include analysis of NVIDIA, the Rubin rack bill of materials (BOM), ADI, CPUs, memory, and semiconductor equipment. Japan and South Korea saw momentum stocks reverse, with capital flowing back towards semiconductors and AI. The South Korean market, a proxy for global tech momentum investing, saw net buying by domestic institutions alongside net selling by foreign and retail investors. China H-shares and A-shares underperformed the broader Asia-Pacific market, although Hong Kong saw net southbound buying.

Two notable charts highlight the Rubin VR200 rack: 1) The quantity of MLCCs increased threefold compared to the GB300, with a value of approximately $4,300 per rack. 2) The value of ABF substrates doubled compared to the GB300, reaching about $20,000 per rack.

**NVIDIA**: Morgan Stanley maintains its Top Pick rating, highlighting that April quarter revenue of $81.6 billion and July guidance of $91 billion both exceeded expectations. The on-schedule delivery of Vera Rubin and a CY26 CPU revenue target of $20 billion are expected to continue suppressing the narrative of ASIC diversification.

**Rubin Rack/ODM Value**: Morgan Stanley estimates the ASP for the VR200 Rubin rack procured from ODMs is approximately $7.8 million, a 95% increase from the GB300. The value of PCBs, MLCCs, ABF substrates, power supplies, and thermal solutions has risen significantly, and ODM absolute profits are also expected to increase.

**Analog Devices**: Bernstein notes that ADI's FQ2 results and FQ3 guidance significantly exceeded expectations, with simultaneous improvement in automotive, industrial, and communications segments. However, the stock price decline reflects growing market concerns about the analog cycle peak, gross margins, and valuations nearing a cyclical high.

**Agentic CPUs**: Bank of America has raised its 2030 server CPU total addressable market (TAM) estimate to $125 billion, up from $110 billion. Agentic AI is expanding the role of CPUs from supporting GPUs to standalone CPU racks.

**Compute Power/Token**: CLSA believes token growth and compute shortages are still intensifying. Analysis of Anthropic's compute structure shows an increasing share for TPUs, but NVIDIA may still recapture incremental share through new orders and its Blackwell/Rubin platforms.

**Memory**: UBS argues that AI agents are expanding memory demand from HBM to include DDR5, LPDDR5, and NAND. Tight supply for DRAM/NAND could persist until around 2028.

**Semiconductors/Equipment**: J.P. Morgan states that companies at the TMC conference collectively confirmed continued strength in orders, backlogs, and customer upgrades. The simultaneous emergence of AI inference demand, memory tightness, and a lengthening Wafer Fab Equipment (WFE) cycle supports maintaining an Overweight rating on the semiconductor and equipment sectors.

**Data Center Thermal Materials**: UBS notes that AI chips with 700W to 1200W power consumption are pushing thermal bottlenecks to the packaging material level. High-performance composite materials can improve thermal conductivity by approximately 3 to 4 times, driving increased demand for both liquid cooling and chip-level first cooling solutions.

**CPO/Silicon Photonics**: Morgan Stanley views the NVIDIA NVL576 and subsequent multi-rack architectures as an acceleration point for Co-Packaged Optics (CPO). The optical engine-to-GPU attachment rate could rise from 2-4 to 17 or even above 35. Price targets for Soitec and Besi have been significantly raised.

**Silicon Capacitors/SEMCO**: Morgan Stanley believes SEMCO's long-term agreement for approximately 1.5 trillion won ($10 billion) in silicon capacitors confirms a larger Si-Cap TAM. This supply is primarily for the next-generation TPU's EMIB-T, providing new structural opportunities for legacy DRAM process suppliers.

**Asian Tech Supply Chain**: J.P. Morgan's Asia Technology Tracker highlights incremental focus on SEMCO, Silergy, Wasion, and Japanese components. AI servers are driving demand for silicon capacitors, ABF substrates, low/medium-voltage Si MOSFETs, and power supply chains.

**Detailed Analysis: NVIDIA - Rubin and CPU Narrative Takes Over, AI Compute Leader Remains Top Pick** Morgan Stanley (May 26, 2021) • Core View: NVIDIA's earnings and guidance comprehensively beat expectations. The on-track progress of Vera Rubin and the CPU revenue target jointly reinforce the "AI factory economics" thesis, making the ASIC diversification narrative difficult to challenge in the near term. • Key Figures: April quarter revenue of $81.615 billion, up 19.8% QoQ and 95.1% YoY; Data Center revenue of $75.249 billion; July quarter revenue guidance midpoint of $91 billion. • Focus Areas: Subsequent focus is on Vera Rubin's Q3 delivery, Computex/GTC disclosures, the $20 billion CY26 CPU target, and the penetration of agentic standalone CPU racks at customers like CoreWeave, META, and Oracle. Morgan Stanley argues that NVIDIA delivered a clean beat and raise despite high expectations. The core takeaway is not the single-quarter beat itself, but how the Vera Rubin production timeline, CPU revenue target, and supply chain commitments collectively support its AI factory economics. The company announced an $80 billion buyback and a $1 annualized dividend. Purchase commitments rose to $119 billion, with $95 billion slated for production over the next three quarters. Valuation remains below AI peer multiples, supporting the maintained Overweight/Top Pick rating.

**Rubin Rack BOM: PCB and ODM Absolute Profits Benefit Most** Morgan Stanley (May 20, 2021) • Core View: Increased complexity, power density, and compute density of the Rubin rack will drive growth in PCB, MLCC, ABF substrate, power supply, thermal solution, and ODM value-added. • Key Figures: VR200 Rubin rack ASP is approximately $7.8 million, a 95% increase from the GB300's ~$3.99 million. PCB value increased +233%, MLCC +182%, ABF +82%, power supply +32%, thermal solution +12%. • Focus Areas: SOCAMM procurement model, consignment business model progress, ODM absolute dollar profit increase, and the ramp-up timing for downstream components like PCB/MLCC/ABF. Morgan Stanley estimates the ASP for the VR200 Rubin NVL72 rack procured from ODMs is about $7.8 million. The cost increase stems not just from GPUs but also from higher memory prices and increased system complexity. Memory's share of the rack BOM rose from ~7-9% in GB200/GB300 to ~26% in VR200, while the GPU share fell from ~65% to ~51%. PCB saw the most significant value increase, estimated at ~$117,000 per rack for VR200 versus $35,000 for GB300. ODM value-added is estimated to rise 35-40% compared to Blackwell racks. While gross margins may compress slightly, absolute dollar profit growth is key. Morgan Stanley's preferred ODM is Wiwynn, followed by Wistron, Quanta, and Hon Hai. Component beneficiaries include Delta Electronics, AVC, Unimicron, Zhen Ding Tech, and FIT.

**Analog Devices: Analog Recovery Continues, But Peak Risks Enter Pricing** Bernstein (May 21, 2021) • Core View: Bernstein believes ADI execution and cyclical recovery are strong, with FQ2 results and FQ3 guidance significantly better than the market expected. However, flattening gross and operating margins amid high growth are introducing peak cycle risks as a valuation constraint. • Key Figures: FQ2 revenue of $3.623 billion, up 15% QoQ and 37% YoY; pro forma EPS of $3.09; gross margin of 73%. FQ3 revenue guidance midpoint of $3.9 billion and EPS of $3.30. Price target raised to $430. • Focus Areas: Subsequent focus is on whether industrial, automotive, and communications can sustain above-seasonal growth, if data center evolves from a small base to a larger driver, and how the Empower acquisition strengthens the silicon capacitor/power portfolio. Despite strong results, the stock fell ~4% after-hours. Bernstein attributes this partly to crowded positioning and investor focus on peaking margins, high utilization rates, and completed pricing actions. The firm raised its FY26-FY28 revenue and EPS estimates and increased the price target from $375 to $430 but maintains a Market-Perform rating.

**Agentic AI Pushes CPUs Back to Data Center Forefront** Bank of America (May 19, 2021) • Core View: Agentic AI is shifting CPU demand from host CPUs within GPU racks to standalone CPU-only racks, leading to an upward revision of the server CPU TAM. • Key Figures: Bank of America raised its 2030 server CPU TAM estimate to $125 billion from $110 billion, implying growth from $43 billion in 2026 and a 2026-2030 CAGR of 31%. • Focus Areas: Key validation points are AMD's continued share gains in x86 cloud/enterprise markets, the ramp of NVIDIA's Vera CPU racks, and increased share for ARM custom/ASIC CPUs. Bank of America emphasizes this is not a story of CPUs replacing GPUs. Agentic AI is system-led. GPUs/XPUs remain the core bottleneck for matrix operations and large model inference. The CPU increment comes from a more balanced system architecture. The standalone Vera CPU rack within the Vera Rubin platform exemplifies this change, where CPU counts could approach GPU counts at the full pod level.

**AI Compute Shortage Intensifies, Semiconductor Valuation Logic Shifts Higher** CLSA (May 19, 2021) • Core View: CLSA believes AI token demand continues to grow faster than expected, and compute supply is unlikely to ease quickly before 2027. The rising share of semiconductors in global tech market cap has fundamental support. • Key Figures: Anthropic has contracted for ~14.46 million chips, with ~1.44 million currently deployed (~10%). Of future incremental chip revenue, Google is estimated at ~50%, NVIDIA ~35%, and AMZN ~15%. • Focus Areas: Key variables for continued re-rating include follow-on Anthropic orders, TPU vs. NVIDIA GPU share shifts, OpenRouter token growth rates, and CoWoS/HBM supply. CLSA states the AI theme has returned to core issues: tokens are exploding, and compute remains scarce. Anthropic is a key case study. While TPUs are changing the accelerator market structure, NVIDIA is not losing opportunities, as uptime and usage limit pressures may drive additional GPU orders. Token demand is the core reason for CLSA's optimism, with agentic tasks consuming far more tokens per request. On the supply side, CLSA estimates 2026 AI accelerator shipments near 14 million units and 2027 near 22 million units, which would largely consume available CoWoS capacity. GPU rental prices are also rising. CLSA upgraded AMD and Marvell to Outperform and remains positive on TSM, NVDA, Broadcom, and AAPL.

**Agentic AI Expands Memory Demand, DDR and NAND Enter New Cycle** UBS (May 20, 2021) • Core View: UBS argues agentic AI is expanding memory demand from HBM to include DDR5, LPDDR5, SOCAMM, and NAND, pushing the memory industry into a longer, more stable new cycle. • Key Figures: UBS forecasts server DDR bit demand to grow 44% YoY in 2026 and 70% in 2027. DRAM tightness is expected to last until at least Q2 2028, and NAND tightness until Q4 2027. • Focus Areas: Key indicators for profit sustainability are the proportion covered by enhanced LTAs, Samsung catching up to SK Hynix in HBM share, and hyperscaler memory spending as a percentage of capex. Agentic AI inference relies more on serial computing and CPU control planes, increasing demand for server CPUs and accompanying DDR5/LPDDR5. NAND is also becoming an incremental driver as KV cache and storage needs rise. UBS expects NAND bit demand to grow 20% in 2026 and 23% in 2027. Enhanced Long-Term Agreements (LTAs) are changing the industry's cyclicality, locking in a significant portion of hyperscaler DDR5 demand and improving cash flow visibility. UBS switched its top pick from SK Hynix to Samsung, raising the target price to 400,000 KRW, citing Samsung's potential to reach ~40% HBM share by 2027 and its capacity to serve multiple customer LTAs.

**Extended Semiconductor Cycle: Inference, Memory, and WFE Strengthen Simultaneously** J.P. Morgan (May 21, 2021) • Core View: Feedback from J.P. Morgan's TMC conference indicates semiconductor and equipment companies continue to see strengthening orders, backlogs, and customer upgrades post-earnings, with AI, memory, WFE, and industrial recovery forming a resonance. • Key Figures: KLA indicated CY27 WFE will be higher than CY26 with the best visibility in over a decade. MKSI is preparing for a $165-180 billion WFE environment post-2027. Micron and Sandisk both point to continued memory tightness. • Focus Areas: Key validation points include CPU:GPU ratio changes from inference workloads, LTA pricing for HBM/DRAM/NAND, CY27 WFE upside, and whether true inventory restocking begins in industrial/automotive/A&D. The most consistent message was that demand momentum continued post-earnings. AI workloads are expanding from pure accelerator demand to a full-stack opportunity. Intel noted the CPU:GPU ratio shifts from ~1:8 in training to ~1:1 in inference, with some agentic workloads reaching 4:1 CPU-heavy. Memory and WFE are hard evidence of a prolonged cycle. Micron's outlook strengthened post-March earnings, and Sandisk extended its supply tightness view to the end of CY27, with several large LTAs already signed. J.P. Morgan's top picks include Broadcom, Marvell, Micron, KLA, Applied Materials, Cadence, Synopsys, Astera Labs, and MKSI.

**Data Center Thermal Solutions: HPA Packaging Materials Address Chip-Level Bottleneck** UBS (May 21, 2021) • Core View: UBS believes AI data center thermal bottlenecks are driving packaging material upgrades. High-Performance Adhesives (HPA) with high thermal conductivity fillers are moving from R&D to market deployment, serving as a chip-level first cooling solution alongside liquid cooling. • Key Figures: 700W-1200W AI accelerators are reaching traditional material limits. HPA EMC thermal conductivity is ~3-4 W/mK versus ~1 W/mK for traditional materials, potentially lowering temperatures by 4-6°C and improving AI throughput by 5-10%. • Focus Areas: Key developments to track include HPA offtake, spherical alumina capacity expansion, NVIDIA and foundry packaging roadmaps, TIM standardization, and the coordination between liquid cooling deployment and high-thermal-conductivity packaging. Recent data center overheating incidents signal tightening thermal design margins. New solutions involve replacing traditional silica filler in packaging layers with high-purity alumina or other ceramic materials for better internal heat spreading. Evidence is shifting from R&D to productization, with several suppliers advancing high-purity, spherical alumina fillers. UBS notes that solving thermal challenges won't rely on a single layer. Second-gen packaging materials provide "first cooling," while chip-level and system-level liquid cooling accelerate in parallel. Widespread adoption of high-thermal-conductivity materials could save 20-30 TWh/year in data center power by 2030.

**CPO Opportunity Reassessed: NVL576 Pushes Silicon Photonics into Scale-Up Networks** Morgan Stanley (May 19, 2021) • Core View: Morgan Stanley believes the NVIDIA NVL576 and subsequent multi-rack AI architectures will push optical communications from scale-out further into scale-up networks, significantly boosting demand for Soitec's Photonics SOI and Besi's hybrid bonding. • Key Figures: The optical engine-to-GPU attachment rate is forecast to rise from the current ~2-4 to 17 for the NVL576 hybrid architecture, 35 for full CPO, and up to 70 in a 200G SerDes scenario. Soitec's price target raised to €200; Besi's to €300. • Focus Areas: Key variables are the NVL576 launch in late 2026 and mainstream adoption by 2028, TSMC COUPE adoption, silicon photonics penetration, and whether HBM4E 16-hi and Feynman GPUs adopt hybrid bonding. As model scale and context windows grow, multi-rack low-latency domains requiring cross-rack data sharing will approach the physical limits of copper, making optical connections a more realistic architectural choice. Morgan Stanley's model breaks down GPU/ASIC assumptions, AI transceiver models, and NVIDIA architecture to estimate optical engine counts. Photonics SOI demand directly benefits Soitec, while TSMC's COUPE technology using SoIC for EIC-PIC stacking supports demand for Besi's hybrid bonding equipment.

**Silicon Capacitor TAM Revised Upward: SEMCO LTA Opens New Uses for Legacy DRAM Processes** Morgan Stanley (May 21, 2021) • Core View: Morgan Stanley believes SEMCO's long-term silicon capacitor agreement confirms an expanding Si-Cap TAM, potentially extending beyond AI servers to HPC, autonomous driving, and mobile devices, with an oligopolistic supply structure driving long-term agreements. • Key Figures: SEMCO signed a silicon capacitor contract worth approximately 1.5 trillion won (~$10 billion), covering the period from January 1, 2027, to December 31, 2028. Morgan Stanley judges this is primarily for the next-generation TPU's EMIB-T. • Focus Areas: Subsequent developments to watch include expansion to new customers like mobile, feedback on AP Memory-related demand, PSMC capacity tightness, and whether Winbond becomes a potential partner for SEMCO. The contract moves silicon capacitors from a niche advanced packaging component toward a clearer long-term TAM. The oligopolistic supply nature makes customers more likely to sign LTAs with core suppliers to ensure power integrity and delivery stability for advanced packaging platforms. On the manufacturing side, legacy DRAM processes are well-suited for Si-Cap fabrication, offering a new avenue for asset reuse for Taiwanese DRAM suppliers. With PSMC capacity full, Winbond is seen as a potential collaborator.

**Asian Tech Supply Chain: AI Servers Continue Diffusing Incremental Demand to Components** J.P. Morgan (May 21, 2021) • Core View: J.P. Morgan's Asia Technology Tracker shows AI infrastructure incremental demand continuing to diffuse to component chains in Korea, Taiwan, Japan, and Hong Kong, benefiting silicon capacitors, server power supplies, ABF substrates, and Si MOSFETs. • Key Figures: SEMCO's price target raised to 1.45 million KRW, with FY26-FY28 EPS estimates raised 5-25%. Silergy's AI ASIC revenue is expected to become more significant in 2H27 and 2028. Wasion's data center revenue could double or triple this year. • Focus Areas: Key subsequent focuses are SEMCO's new business execution, Silergy's breakthrough with a US tier-1 CSP, the outcome of Samsung's labor union vote, Wasion's AIDC backlog, and whether supply tightness for Japanese low/medium-voltage Si MOSFETs increases further. Samsung Electro-Mechanics benefits from GPU/CPU orders, mix improvement, and pricing power with long-tail customers, while its silicon capacitor and fabless business model offer new growth vectors. Silergy recently qualified with a US tier-1 CSP, a key milestone for entering the AI server market. Samsung Electronics management reached an agreement with its union, likely removing the strike overhang. Japanese components also point to an AI migration, with ABF substrate production remaining strong and low/medium-voltage Si MOSFET supply tightening due to AI data center demand. Wasion benefits from potential tripling of data center revenue this year and clear backlog visibility for 2028-2029.